Optimization Of Alarm Threshold For Key Vibration Parameters Of Hums System Of A Helicopter

Along with the vigorous development of general aviation,the health management of helicopters has become increasingly important.The HUMS of a certain model of helicopter in service takes fault threshold as the basis for maintenance work.But the actual working conditions are very complex.There are situations where the vibration eigenvalues exceed the fault threshold,but no maintenance is needed,i.e.false maintenance alarm.Addressing this exact problem,this paper proposes adopting the maintenance threshold to enhance the accuracy of HUMS recognition of maintenance,for the purpose of reducing false maintenance alarm and optimizing the alarming threshold of the HUMS.Major failure models and types of accessories monitored by the HUMS and feature parameters of the system are explained.According to the failure probability in the accessories,key modules for fault detection are determined,whose vibration characteristics are then extracted from the system.Based on analysis of the distribution of vibration samples,explanation of the necessity of sample partition and division of the threshold available cycles,a sample partition plan combining anomaly detection with sliding window is put forward so as to include samples of similar intensity of anomaly into the same threshold available cycle.Methods for anomaly detection are also given,including:the statistical learning-based method,the similarity-based method,and the ensemble learning-based method.The principle and use of sliding windows are introduced,too.Results of the partition are compared to determine the appropriate partition plan and eventually obtain all threshold available cycles.By thresholding within a single threshold available cycle and piecing the corresponding threshold of each cycle together in order,we have the threshold curve.Here,three thresholding models are described in detail,including:the model combining Gaussian Mixture Model with nonparametric approach,the block extreme model,and the Chebyshev model given by the system designer.Then,the performance of the three models is evaluated by using the confusion matrix and F_βscore.Based on the evaluation results,the application scenarios of the models are described.In the end,the threshold prediction model based on sample similarity is proposed.